What are prompts?

The basics

Every time you type something into ChatGPT, Claude, Gemini, or any AI tool and hit send, that message is a prompt. The AI reads the prompt, predicts a response, and sends it back. Simple.

The reason "prompts" got a fancy name is that the way you phrase your request has a huge effect on what comes back. A vague prompt produces a vague answer. A specific prompt with the right context produces something you can actually use.

The five things every good prompt has

You don't need all five every time. But for anything important, run through this list.

• Role. Who you want the AI to be. "Act as a real estate lawyer." "You are a small business marketing consultant." This sets the lens.
• Task. What you want it to do. "Review this lease." "Write a follow-up email." Be specific about the action.
• Context. What it needs to know. Who's the audience? What's the situation? What's already been tried? This is the most-skipped part and the one that matters most.
• Format. What the output should look like. "Bullet points." "Three paragraphs." "A table with these columns." "A draft email I can send."
• Constraints. What to avoid or include. "Under 200 words." "No legal jargon." "Use plain English." "Don't include disclaimers."

Things that work surprisingly well

• Give examples. "Here are two emails I've written before. Match this voice." Two or three examples beats a paragraph of style rules.
• Tell it what good looks like. "A great answer would include X, Y, and Z." This dramatically improves output.
• Ask it to think first. For complex tasks, "before you answer, walk through your reasoning step by step." You often get a much better final answer.
• Let it ask questions. "If anything is unclear, ask me before you start." Stops AI from making bad assumptions.
• Tell it the audience. "Explain this to a 14-year-old." "Write this for a partner-track lawyer." Same content, totally different output.

Things that don't work as well as people claim

• "Act like a Harvard professor." Pretentious role-play rarely helps. Specific roles do. "Act as a contracts paralegal at a 50-attorney firm" is better than "act as the world's smartest lawyer."
• "I'll tip you $200." Used to be a meme. Doesn't move the needle on modern models.
• Super long prompts. Past about 500 words, you start confusing the model. Structure beats length.
• All caps and exclamation marks. Doesn't make the model take you more seriously. It just makes your prompt harder to parse.

The intermediate move: iterate, don't perfect

Don't try to write the perfect prompt in your head before typing. Send a decent first version, see what you get, then refine. Most professionals work like this:

1. Send a quick first prompt.
2. Look at the output. What's off?
3. Reply with corrections. "Make it shorter. Drop the second point. Use a more casual tone."
4. Repeat until it's right.

The chat is a conversation, not a search query. Treat it that way.

One-shot vs few-shot prompts

You will see these terms. They are simpler than they sound.

• Zero-shot. You just ask. "Write a product description for this." No examples.
• One-shot. You give one example. "Like this one we wrote for our last product. Now do one for this new one."
• Few-shot. You give 2 to 5 examples. The model picks up the pattern and matches it.

Few-shot is the secret weapon for anything with a "house style." Customer support replies, product descriptions, social posts, intake summaries. Show the AI three examples and it will copy the pattern almost perfectly.

The advanced take: prompts vs instructions vs system prompts

Strictly speaking, what most people call a "prompt" is just the user message. But there are layers:

• System prompt. Hidden setup that the AI sees before any conversation. Tools like ChatGPT and Claude have one even when you don't.
• Custom instructions. Your personal standing rules that get attached to every chat (see topic 06).
• User prompt. What you actually type.

The AI reads all of them together. So if you set up custom instructions well, your day-to-day prompts can stay short because the heavy context is already loaded.